Speech-processing apparatus and speech-processing method

Abstract

A speech-processing apparatus includes: a representative transfer function estimation unit that uses a sound signal which is collected by using a microphone array of which the arrangement is unknown, which has a plurality of channels, and of which the number of sound sources is unknown and that estimates a transfer function with respect to a sound source.

The invention claimed is: 
     
       1. A speech-processing apparatus, comprising:
 a processor coupled to a memory and configured to: 
 use a sound signal which is collected by using a microphone array of which the arrangement is unknown, which has a plurality of channels, and of which a number of sound sources is unknown and estimate at least one transfer function with respect to a sound source; 
 detect a plurality of speech zones from the sound signal that has the plurality of channels; 
 estimate the at least one transfer function by using a speech signal of a detected speech zone; 
 calculate a feature quantity by using the speech signal; 
 estimate the number of sound sources by classifying the calculated feature quantity into clusters, wherein each transfer function to each one of the clusters is estimated as each transfer function with respect to a sound source based on the estimated transfer function and the estimated number of sound sources; and 
 store data representing recognition results for the number of estimated sound sources on the memory. 
 
     
     
       2. The speech-processing apparatus according to  claim 1 , wherein
 one from the plurality of channels included in the speech signal is selected as a representative channel, and the at least one transfer function is estimated by using the speech signal of the selected representative channel and the speech signal of the rest of the plurality of channels. 
 
     
     
       3. The speech-processing apparatus according to  claim 1 , wherein
 one from the plurality of channels included in the speech signal is selected as a representative channel, and the feature quantity is calculated by using the speech signal of the selected representative channel and the speech signal of the rest of the plurality of channels. 
 
     
     
       4. The speech-processing apparatus according to  claim 3 , wherein
 the feature quantity is a value based on a phase difference between the speech signal of the representative channel and the speech signal of the rest of the plurality of channels. 
 
     
     
       5. The speech-processing apparatus according to  claim 1 , wherein
 a number of the clusters is determined based on a distance between the clusters to which the number of integrated clusters converges. 
 
     
     
       6. The speech-processing apparatus according to  claim 1 , wherein
 a center of each one of the obtained clusters is calculated, 
 with respect to the feature quantity included in each cluster, a distance between the feature quantity and the calculated center of each one of the clusters is calculated, 
 a weighted value is calculated by using a sum of the calculated distances and each one of the calculated distances, and 
 by using a transfer function corresponding to the feature quantity included in the cluster and the calculated weighted value, the transfer function in the center of the cluster is estimated as the transfer function to a sound source. 
 
     
     
       7. A speech-processing method that causes a processor coupled to a memory to:
 (a) acquire a sound signal which is collected by using a microphone array of which the arrangement is unknown, which has a plurality of channels, and of which the number of sound sources is unknown; 
 (b) detect a plurality of speech zones from the sound signal that has the plurality of channels; 
 (c) estimate a transfer function by using a speech signal of a speech zone detected in (b); 
 (d) calculate a feature quantity by using the speech signal; 
 (e) estimate the number of sound sources by classifying the feature quantity calculated in (d) into clusters; and 
 (f) estimate each transfer function to each one of the clusters as each transfer function with respect to a sound source based on the transfer function estimated in (c) and the number of sound sources estimated in (e). 
 
     
     
       8. A speech-processing apparatus, comprising:
 a sound signal acquisition unit configured to acquire a plurality of sound signals having a plurality of channels; 
 a speech estimation unit configured to extract a sound signal of one session from the plurality of sound signals; 
 a conversion unit configured to output a speech signal for a speech zone of the number of speech zones based on the sound signal; 
 a transfer estimation unit configured to generate transfer functions for the speech zone of the number of speech zones; 
 a representative transfer function estimation unit configured calculate a feature quantity from the converted speech signal and estimate a number of sound sources by classifying the feature quantity into clusters, wherein each transfer function is applied to each one of the clusters to estimate a sound source of a plurality of sound sources; and 
 an output unit configured to output data representing recognition results for the estimated sound source. 
 
     
     
       9. The speech-processing apparatus of  claim 8 , further comprising:
 a speech zone estimation unit configured to detect the number of speech zones and outputs a number a speech signals for each detected speech zone. 
 
     
     
       10. The speech-processing apparatus of  claim 9 , wherein the speech zone estimation unit detects the number of speech zones using a zero-cross method. 
     
     
       11. The speech-processing apparatus of  claim 8 , wherein the conversion unit converts the speech signal using a short-time Fourier transform into a frequency domain. 
     
     
       12. The speech-processing apparatus according to  claim 8 , wherein a number of the clusters is determined based on a distance between the clusters to which the number of integrated clusters converges. 
     
     
       13. The speech-processing apparatus according to  claim 8 , wherein
 a center of each one of the obtained clusters is calculated, 
 with respect to a feature quantity included in each cluster, a distance between the feature quantity and the calculated center of each one of the clusters is calculated, 
 a weighted value is calculated by using a sum of the calculated distances and each one of the calculated distances, and 
 by using a transfer function corresponding to the feature quantity included in the cluster and the calculated weighted value, a transfer function in the center of the cluster is estimated as a transfer function to a sound source.